Brian Geisel, CEO of Geisel Software, presents the “Using Synthetic Data to Train Computer Vision Models” tutorial at the May 2024 Embedded Vision Summit.
Developers of machine-learning based computer vision applications often face difficulties obtaining sufficient data for training and evaluating models. In this talk, Geisel explores the use of synthetic data techniques to overcome these challenges. He explains the “simulation-to-reality” gap and the challenges it poses for realistic synthetic data generation.
Using an automated Mars exploration vehicle application, Geisel shares how his company generates synthetic data using the Unity simulation environment, and demonstrates how synthetic data can be effective—while also highlighting its limitations. He also touches on the potential for few-shot learning techniques to reduce the need for training data.
See here for a PDF of the slides.